Hippocampal neurons encode word meanings through distributed, LLM-like population codes

The human hippocampus represents word meanings not through single “concept cells” but through distributed patterns of activity across many neurons—a coding scheme that mirrors the semantic geometry of large language models (LLMs) like GPT-2. The finding, published today in Nature Neuroscience, offers a new neurocomputational account of how the brain extracts meaning from natural speech.

The work, led by neuroscientist Benjamin Y. Hayden at Baylor College of Medicine, with first author Melissa Franch, recorded 356 hippocampal neurons in ten epilepsy patients while they listened to narrative podcasts. Using encoding models, the team found that semantic features predicted neural firing even after controlling for phonetics, syntax, and other linguistic cues. On average, each word was encoded by 37 neurons, and individual neurons responded to words spanning many unrelated semantic categories—a hallmark of mixed selectivity.

Strikingly, the population distance between neural responses to two words tracked their semantic distance as defined by contextual embeddings from GPT-2 and BERT, but not by static embeddings like Word2Vec. The neural code also captured polysemy: the more meanings a word had, the more variable its neural representation across different contexts. For very similar words, the researchers observed a contrastive effect—neural representations were pushed further apart than embeddings would predict—which may help the brain avoid confusing highly confusable meanings.

The study is observational and based on a limited number of patients, and the effect sizes, while robust, are modest—expected given the single-presentation, naturalistic design. The authors caution that the LLM parallel reflects shared geometric principles, not that the brain operates like a transformer.

Still, the results are a significant step toward understanding how the hippocampus, a structure best known for memory, contributes to real-time language comprehension. Future work may extend these single-neuron recordings to other brain regions to see how widespread this distributed, contextual coding strategy is.

Reference: Franch, M., Mickiewicz, E.A., Belanger, J.L. et al. A population code for semantics in human hippocampus. Nat Neurosci (2026). https://doi.org/10.1038/s41593-026-02436-4